app.py

# app.py — Your original working version + repetition_penalty=1.2 + ngram=3

import os
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"

import gradio as gr
import spaces
from transformers import pipeline
import torch
import gc

# ——————————————————————————————
# ZeroGPU worker – model loaded inside
# ——————————————————————————————
@spaces.GPU(duration=180)
def transcribe_3min(audio_path):
    if not audio_path:
        return "Hlaðið upp hljóðskrá"

    pipe = pipeline(
        "automatic-speech-recognition", 
        #model="palli23/whisper-tiny-icelandic-distilled-v3",
        #model = "palli23/whisper-tiny-distilled-spjallromur-polish-v3",
        #model = "palli23/whisper-tiny-distilled-spjallromur-polish-v5",
        #model="palli23/whisper-tiny-distilled-samromur-spjallromur-polish",
        #model="palli23/whisper-tiny-samromur-spjallromur",
        model="palli23/whisper-small-sam_spjall",
        torch_dtype=torch.float16,
        device=0,  # GPU inside @spaces.GPU
    )

    result = pipe(
        audio_path,
        chunk_length_s=30,
        batch_size=8,
        return_timestamps=False,  # ← no timestamps, as you want
        generate_kwargs={
            "num_beams": 5,  #var beam size 1
            "repetition_penalty": 1.2,     # ← exactly what you asked for
            "no_repeat_ngram_size": 3,     # ← exactly what you asked for
            "temperature": 0.0,
        }
    )

    # Clean memory so ZeroGPU lives forever
    del pipe
    gc.collect()
    torch.cuda.empty_cache()

    return result["text"]

# ——————————————————————————————
# UI – clean and simple
# ——————————————————————————————
with gr.Blocks() as demo:
    gr.Markdown("# Íslenskt ASR – 3 mínútur")
    gr.Markdown("**palli23/whisper-small-sam_spjall** · mjög lágur WER · allt að 5 mín hljóð")
    gr.Markdown("**Hafa samband:** pallinr1@protonmail.com")

    audio_in = gr.Audio(type="filepath", label="Hlaðið upp .mp3 / .wav")
    btn = gr.Button("Transcribe", variant="primary", size="lg")
    output = gr.Textbox(lines=25, label="Útskrift")

    btn.click(fn=transcribe_3min, inputs=audio_in, outputs=output)

# ——————————————————————————————
# Public launch
# ——————————————————————————————
demo.launch(
    share=True,
    server_name="0.0.0.0",
    server_port=7860,
    auth=None
)